Image analysis and mathematical methods for diagnosis in pathology: Applications to neoplastic diseases of the breast, ovary, and parathyroid.
Item
-
Title
-
Image analysis and mathematical methods for diagnosis in pathology: Applications to neoplastic diseases of the breast, ovary, and parathyroid.
-
Identifier
-
AAI9820528
-
identifier
-
9820528
-
Creator
-
Einstein, Andrew Jeffrey.
-
Contributor
-
Adviser: Joan Gil
-
Date
-
1998
-
Language
-
English
-
Publisher
-
City University of New York.
-
Subject
-
Health Sciences, Medicine and Surgery | Mathematics | Health Sciences, Pathology | Health Sciences, Oncology
-
Abstract
-
Image analysis and mathematical tools are developed for the resolution of difficult differential diagnoses in pathology. Methodological developments center on three main areas: quality assurance, the quantitative characterization of chromatin appearance, and classification. The impact of a new cytologic staining method, the ultrafast Papanicolaou procedure, on nuclear morphometry is evaluated. It is found to improve on conventional Papanicolaou staining in terms of speed, with no important quantifiable differences in nuclear morphology. Variance component models are used to analyze the reproducibility of several interactive methods used for the segmentation of nuclear images, demonstrating that thresholding-based methods similar to those incorporated into commercial instruments are unacceptably variable, while an arc-fitting method is most reproducible. Two new approaches to chromatin texture characterization are introduced, a statistical method quantitating the amount of local gray level variation, and an approach based on fractal geometry. Chromatin appearance in breast epithelial cell nuclei is shown to be fractal. Nuclei are characterized by Minkowski-Bouligand and spectral fractal dimensions, and by lacunarity, a measure of a fractal's "gappiness." These properties are shown to differ between benign and malignant breast epithelial cell nuclei. Classification methods are compared, and logistic regression and artificial neural network models are developed and optimized.;These methods are applied to the cytologic diagnosis of breast epithelial cell lesions, and histopathologic diagnoses of parathyroid lesions and ovarian dysplasia. Using fractal measures of chromatin appearance, and logistic regression for classification, 39 of 41 breast cytology specimens are correctly diagnosed as benign or malignant. An optimized neural network is able to correctly classify all 41 cases; however, questions of overtraining are raised. Using nuclear diffuseness, the measure of gray level variability, to characterize chromatin appearance, and a neural network for classification, all 16 parathyroid biopsies considered are correctly diagnosed as normal, adenoma, or carcinoma. Statistical and neural network analyses show ovarian dysplasia in epithelial inclusion cysts adjacent to cancer and incidentally diagnosed to be indistinguishable. This suggests that the associated changes in tissue morphology reflect a single underlying pathologic process, supporting the hypothesis that ovarian dysplasia seen in epithelial inclusion cysts represents a preinvasive malignant change.
-
Type
-
dissertation
-
Source
-
PQT Legacy CUNY.xlsx
-
degree
-
Ph.D.
